More details

Count distinct complainants and normalize exposure

C24 is the number of source-scoped authors with at least one accepted complaint in the preceding 24 hours. C2 counts authors whose first accepted complaint in that window appeared during the latest two hours. Repeated posts by the same author do not add another complainant.

P is the actual number of posts checked. The nominal exposures are 1,200 posts for 24 hours and 100 posts for two hours.

Build the comparable baseline

The baseline uses at least 28 earlier complete windows ending at the same UTC hour. A window d days old receives relative weight 2^(-d/21), so its influence halves every 21 days. Incomplete, incompatible, and pre-excluded reviewed incident windows do not enter the baseline.

The predictive variance uses the weighted spread plus count overdispersion. Its floor is 1.0 for 24 hours and 0.5 for two hours. If the current sample is thinner than its weighted history, the detector increases variance and therefore lowers confidence.

Calculate standardized anomalies and the final score

Each normalized count is measured against its baseline in standard-deviation units. The two-hour term can add only a positive, capped boost. If that component is unavailable, its boost is zero.

The displayed ratio is (C24 normalized + 0.5) / (baseline expectation + 0.5). It is context only and does not set the signal. The displayed expected count is converted back to the current observed-post exposure.

Apply thresholds and evidence safeguards

Any public state requires complete evidence and at least 400 sampled posts in 24 hours. Probable issue and Strong signal require at least 600 posts and evidence beyond one similar-text cluster; otherwise the result can be no higher than Watch.